Page 362 - System on Package_ Miniaturization of the Entire System

334 Cha pte r S i x

edge-viewing photodetectors are used. The term “self-aligning” means that the active
devices and passive lightwave circuits are all simultaneously aligned during the process
of waveguide core path definition, accomplished by selectively cross-linking monomers
either by direct UV laser writing, projection lithography, or proximity lithography
through a mask. The optical wiring density is limited primarily by the dimensions of
laser bare dies and photodetector bare dies. To overcome this potential limitation, the
actives can be staggered in a two-dimensional array, and waveguide escape routes can
be designed to converge in a parallel array. It should be noted that, because optical
interconnects can transcend electrical wiring hierarchy, board-to-board optical wiring
can originate and terminate on any board edge or anywhere on the board front or back
surfaces, as compared to electrical wiring which is defined at a board’s edge. For optical
interconnects, it is simpler to refer to the total number of interboard optical links, which
are separate from the total number of intraboard optical links.

6.4.3 Power Dissipation
Simulations of the power dissipation by optical and electrical interconnects, on-chip
and on-board, inevitably compare one copper line with one optical link. The one-for-
one comparison is not realistic. The task of comparing power dissipation between
electrical and optical interconnects has to take the following into account:
1. The electrical bus aggregate bit carrying capacity: width, length, and bit rate
per line. This should be compared with the number, usually much lower, of
optical lines that are needed for the same aggregate bandwidth.
2. The aggregate CMOS power dissipated in driving the electrical bus, including
line equalization, multiplexers, and demultiplexers, versus the CMOS or
bipolar power dissipated in driving the (fewer) number of laser drivers and PD
amplifiers that replace the electrical bus.

The determining factors are distance and aggregate bandwidth. The type of laser used
is a perturbation effect. Two independent authors have shown that when all components
in the optical link (laser drivers and photodetector amplifiers) are taken into
consideration, there is little difference in the power consumption of optical links based
on VCSELs or based on edge-emitting lasers (EELs) [20, 27–29]. The important difference
between VCSELs and EELs is the fact that readily available commercial VCSELs roll
off rapidly after 3.3 Gb/s [30], while readily available commercial EELs roll off after
10 Gb/s. On the horizon, Agilent and Emcore are developing 20-Gb/s VCSELs, in
part with DARPA funding, while NTT has reported operation of a directly modulated
40-Gb/s EEL [31].
Receiver equalization can extend the range of a 10-Gb/s, differential microstrip
transmission line to about 0.75 m on an FR4 or PCL-FR3 [32] substrate using non-return-
to-zero (NRZ) serial bit encoding. The equalization method is deterministic and
compensates mostly for intersymbol interference (ISI) phase shift noise due to frequency-
dependent skin effect and dielectric losses. For example, the MAX3804 receiver equalizer
is designed to do just that and can significantly decrease deterministic jitter. The Maxim
MAX3804 has a footprint of 3 mm × 3 mm and dissipates 115 mW per differential copper
line [33].
A 10 Gb/s per channel, 1 × 12 VCSEL driver from Helix AG dissipates 120 mW per
channel, and a 10 Gb/s, 1 × 12 photodetector amplifier, also from Helix AG, dissipates